Results

Main Activities in Preparation and Execution

The PoC execution was organized into several sprints, focusing on iterative development and deep collaboration.

Data Acquisition and Preparation

Data Ingestion: The first step involved ingesting diverse data streams required for the two Use Cases, demonstrating the non-intrusive integration capability of the service.

  • DSO Data: 15-minute net-load time series from EPR’s AMI Head-End System (HES) for customers in the self-consumption scheme. PV site metadata (location, capacity) from the EPR database.
  • EO Data: Hourly Copernicus ERA5 reanalysis and EUMETSAT Meteosat Second Generation (MSG) data (irradiance, cloud properties) were sourced to provide high-resolution spatial coverage.
  • Reference Data: Clean PV production data was obtained from a limited number of sub-metered systems (e.g. Reduxi-equipped customers) and larger PV installations
    > 40kW to serve as ground truth for model validation.

Data Processing: An Extract, Transform, Load (ETL) pipeline was implemented for cleaning, temporal synchronization, and spatial alignment of all input data, which is crucial before model training.

Model Development and Testing

Multi-Model Approach: For each Use Case, ML models were developed and tested using various combinations of inputs to quantify the added value of the Earth Observation (EO) data:

  • Baseline Model: Used only historical patterns and asset metadata.
  • NWP Model: Added Numerical Weather Prediction data.
  • EO Model: Added satellite-derived features (cloud mask, irradiance).
  • Best Configuration: Combined EO and NWP features (with reference PV data, where available).

Iterative Validation: Performance (measured by wRMSE/wMAE, Weighted Root Mean Squared Error / Weighted Mean Absolute Error) was iteratively benchmarked against available ground-truth data (Reduxi PV production and sub-metered net consumption, Reduxi data has been used for this purpose) to ensure the models reached operationally useful accuracy.

Co-Development and Operational Validation

Workshops with Elektro Primorska: Multiple workshops were held with EPR’s engineers (Network Planning and Operations) to define the specific technical requirements, align the service outputs with their planning (UC1) and operational (UC2) workflows, and confirm the business impact.

Infrastructure Fit: Validation confirmed that the EO x Grid modules can communicate with EPR’s systems via existing export/API mechanisms, integrating non-intrusively without changes to their core SCADA or AMI systems.

Key KPIs and their validation

Economic KPIQuantification based on EPR data (2025–2034 Plan)
CAPEX Deferral (UC1)Optimisation of 1–3% of annual investment in LV/MV substations and lines.
Reduced Technical Losses (UC2)4–6% reduction in LV grid technical losses.